In crodriguez-saltos/hummingbrain:

By using a combination of functions from seewave and hummingbrain, one can extract trajectories of several acoustic features as well as summary statistics for those trajectories. This document shows examples of how to obtain these data. For all the examples, we will work with a mashup of two sounds, orni and tico from seewave. Between them, we get good diversity of spectral properties; while orni is a broadband sound, tico is narroband with frequency modulation.

library(seewave)
data(orni)
data(tico)

sound <- pastew(orni, tico, output= "Wave")
spectro(sound, scale= F)

The acoustic measures that we will obtain are slightly modified from those of @barker2008bird.

Trajectories

Sound as input

The following analyses can be done directly with a Wave object as input

Number of elements

This number can be easily obtained with the following code.

library(misound)
elements <- detect_events(wave= sound, msmooth= c(512,0), threshold= 5)
nrow(elements)

Duration of song

Duration of song can be obtained from the elements object. It will simply be the end timestamp of the last element minus the start timestamp of the first syllable.

songdur <- elements$end[nrow(elements)] - elements$start[1]
songdur

Dominant frequency

dfreqs <- dfreq(sound, wl = 512, ovlp = 0, threshold = 5)

Fundamental frequency

funds <- fund(sound, wl= 512, ovlp = 0, threshold= 5)

Spectrogram as input

For the following analyses we need to first calculate a spectrogram. We will do so with the following code.

library(misound)
spectrog <- signal_spectro(
  sound, plot= T, 
  segment_params = list(
    wl= 512, ovlp= 0, threshold= 5
  )
)

Many of the acoustic features of interest in this section can be obtained by a single function, specprop from seewave. We will apply this function across all spectral windows of our spectrogram and store it in a data frame, then, in each subsection we will extract the variable of interest from that data frame.

library(plyr)
specprops <- apply(
  X = spectrog$amp, 
  MARGIN = 2, 
  FUN = function(x){
    if (!all(is.na(x))){
      specprop(cbind(spectrog$freq, x))
    }else{
      NA
    }
  })

# Coalesce numbers into a data frame
NApos <- sapply(specprops, function(x) all(is.na(x)))
specpr <- ldply(.data = specprops[!NApos], 
                   .fun = function(x) as.data.frame(x))
specpr$pos <- which(!NApos)
addnas <- matrix(nrow = length(which(NApos)), ncol = ncol(specpr) - 1)
addnas <- cbind(addnas, which(NApos))
addnas <- as.data.frame(addnas)
names(addnas) <- names(specpr)
specpr <- rbind(specpr, addnas)
specpr <- specpr[order(specpr$pos),]

rm(addnas, specprops, NApos)

summary(specpr)

IQR

The interquartile range is related to the energy splitting difference from @laiolo2003evolution, which was cited by @barker2008bird. That measure is just the average of the IQR calculated at the beginning, middle, and end of the spectrogram. Here, we will obtain the IQR for all spectral windows as a trajectory.

iqrs <- specpr$IQR
plot(spectrog$time, iqrs)

Average frequency

meanfreqs <- specpr$mean
plot(spectrog$time, meanfreqs)

Minimum and maximum frequency

Getting the minimum and maximum frequency is tricky. A fast way of doing so is by establishing a signal/noise threshold for the frequency axis of the spectrogram. So far, we have been using 5% of the maximum amplitude as the threshold for the temporal component of the sound. We will test, in a 3D plot of the spectrogram, whether that same value makes sense as a threshold of the frequency component.

library(plotly)

amplitudes <- na.omit(as.numeric(spectrog$amp))
ampt <- (5 * (max(amplitudes) - min(amplitudes)) / 100) + min(amplitudes)

plot_ly(z= spectrog$amp)%>%
  add_surface()%>%
  add_surface(
    z= matrix(ampt, 
                 ncol= ncol(spectrog$amp), 
                 nrow= nrow(spectrog$amp))
  )

As seen above, a treshold of 5% seems to be doing a good job at separating signal from noise in the frequency axis. We will use this threshold to find maximum and minimum frequency.

library(ggplot2)
library(reshape2)
signalamp <- spectrog$amp > ampt
signalamp <- apply(signalamp, 2, function(x){
  which(x)
})
maxfreqs <- sapply(signalamp, function(x){
  if (length(x) == 0){
    NA
  }else{
    spectrog$freq[max(x)]
  }
})
minfreqs <- sapply(signalamp, function(x){
  if (length(x) == 0){
    NA
  }else{
    spectrog$freq[min(x)]
  }
})

maxmin <- data.frame(
  time= spectrog$time, 
  minfreq= minfreqs,
  maxfreq= maxfreqs
)
maxmin <- melt(maxmin, id.vars= "time", variable.name = "limit")

p <- ggplot(data = maxmin, mapping = aes(x= time, y= value, color= limit)) +
  geom_point()
print(p)

Frequency range

This is the difference between maximum and minimum frequency.

freqrange <- maxfreqs - minfreqs
plot(spectrog$time, freqrange)

Additional metrics

The following metrics were not in @barker2008bird. They can all be obtained from the data frame generated after using specpropr.

Entropy

entropies <- specpr$sh
plot(spectrog$time, entropies)

Median

medians <- specpr$median
plot(spectrog$time, medians)

IQ25

iq25 <- specpr$Q25
plot(spectrog$time, iq25)

IQ75

iq75 <- specpr$Q75
plot(spectrog$time, iq75)

Summary statistics across entire sound

Using trajectories as input

Frequency range

range(maxmin$value, na.rm= T)

Fundamental frequency

It does not make sense to calculate a fundamental frequency for the entire sound. A better approach is to get summary statistics of the trajectory of fundamental frequencies.

summary(na.omit(funds[,2]))

Using mean power spectra as input

meansp <- meanspec(sound, wl= 512, ovlp = 0)

General properties with specprop.

specpr.spec <- specprop(meansp)

Dominant frequency

This is the mode of the power spectra.

specpr.spec$mode

IQR

specpr.spec$IQR

Average frequency

specpr.spec$mean

Median frequency

specpr.spec$median

Entropy

specpr.spec$sh

Q25

specpr.spec$Q25

Q75

specpr.spec$Q75

Run all analyzes as a bundle

The function audioSummary2020, included in hummingbrain, runs all the analyzes from above for any given wave object. We will try it in the next chunk.

library(hummingbrain)
audio_summary <- audioSummary2020(
  wave= sound,
  threshold= 5,
  window_length = 512, 
  overlap = 0, 
  plot= T
)

References

crodriguez-saltos/hummingbrain documentation built on Feb. 1, 2020, 3:07 a.m.